2116 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL…roman/professional/... ·...

2116 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 60, NO. 8, AUGUST 2013

CMOS Tunable-Color Image Sensor With Dual-ADCShot-Noise-Aware Dynamic Range Extension

Derek Ho, Student Member, IEEE, M. Omair Noor, Ulrich J. Krull, Glenn Gulak, Senior Member, IEEE, andRoman Genov, Senior Member, IEEE

Abstract—A wide dynamic range CMOS tunable-color imagesensor is presented. The sensor integrates an 8x8 array of tunable-color photogates which exploit the wavelength-dependent opticalabsorption properties of the polysilicon gate structure. An anal-ysis is presented for the wide dynamic range asynchronous self-reset with residue readout architecture where photon shot noiseis taken into consideration. An implementation of this architec-ture is presented where the (coarse) asynchronous self-reset op-eration and (fine) residue analog-to-digital conversions are per-formed with separate in-pixel and off-pixel circuits, respectively,for a noise-optimized design. A prototype was fabricated in a stan-dard 0.35 µm CMOS process and is validated in color light sensingwhich achieves SNRs of 24.3 dB and 28.5 dB in green and red lightmeasurements, respectively, under a moderate input light inten-sity of 300 µW/cm². The readout circuit achieves a measured dy-namic range of 82 dB with a peak SNR of 46.2 dB under broad-band illumination. The prototype has been integratedwith amicro-fluidic device and experimentally validated in fluorescence contactimaging.

Index Terms—CMOS image sensor, fluorescence imaging,lab-on-a-chip, microfluidics, quantum dot, self-reset, subrangingarchitecture, two-step ADC, wide dynamic range.

I. INTRODUCTION

F LUORESCENCE-BASED transduction is an establishedtechnology and finds a multitude of applications in the

life sciences. For many analytes, it provides the highest sensi-tivity and selectivity from common transduction methods [1].In particular, laser-induced fluorescence is a prominent sensorymethod for lab-on-a-chip devices, depicted in Fig. 1 [2]. Sev-eral groups have focused on the development of integrated fluo-rescence-based sensing platforms for applications ranging fromcancer research [3], [4] to nucleic acid detection [1], [5].One of the advantages of fluorescence-based sensing is its

suitability for spectral multiplexing. To concurrently analyzemultiple biological processes (e.g., hybridization in DNA anal-ysis [6]) or biological structures (e.g., internal organs in smallanimal imaging [7]), multiple fluorescent markers can be used,

Manuscript received August 03, 2012; revised October 24, 2012; acceptedNovember 08, 2012. Date of publication January 29, 2013; date of currentversion July 24, 2013. This paper was recommended by Associate EditorR. Carmona Galan.D. Ho, G. Gulak, and R. Genov are with the Department of Electrical and

Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada(e-mail: [email protected]).M. O. Noor and U. J. Krull are with the Department of Chemical and Phys-

ical Sciences, University of Toronto Mississauga, Mississauga, ON L5L 1C6,Canada.Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TCSI.2013.2239115

Fig. 1. Schematic cross-section of a CMOS fluorescence contact imagingmicrosystem.

which can be distinguished by their emission wavelengths (i.e.,color).Fluorescent markers, such as the green-emitting and

red-emitting quantum dots (QDs) [8], absorb narrow-bandexcitation light and emit at longer wavelengths. Emissionwavelengths are typically between 500 nm to 700 nm andwell-separated (e.g., 50–100 nm) amongst each other. Unlikeother spectroscopic applications where continuous fine resolu-tion spectroscopy techniques (e.g., Raman spectroscopy) arerequired, fluorescence imaging requires spectral differentiationamong only a small number of discrete wavelengths. Therefore,a tunable-color light sensor that can detect a finite number ofwell-separated wavelengths can be used to sense and differen-tiate the emission wavelengths of various fluorophores.In fluorescence imaging, fluorophore-labeled biological sam-

ples can vary widely in the amount of light they output. Forexample, in hybridization assays, target analyte concentrationsin the order of nano- to milli-molar are typical [2], [9]. In ad-dition, the fluorescence excitation light intensity is typically or-ders-of-magnitude higher than that of the fluorescence emission.Coupled with the fact that it is difficult to fabricate ultra-thin yethigh-performance optical filters, the detection of fluorescencein a contact imaging microsystem may have to be performed inthe presence of inadequately-rejected stray excitation light [5].Therefore, employing a wide dynamic range (WDR) imager isadvantageous as it allows for sensing of the low target specimenconcentrations superimposed on a substantial background [10].The choices of the photodetector for fluorescence imaging

systems have conventionally been the photo multiplier tube(PMT) and the charge-coupled device (CCD). PMTs areamongst the most sensitive photodetectors, but are bulky,expensive and require high operation voltage making themunattractive to be integrated into a miniaturized system. Thethroughput of PMT-based detection systems is relatively lowdue to the lack of parallelism in a single photodetector based

1549-8328/$31.00 © 2013 IEEE

HO et al.: CMOS TUNABLE-COLOR IMAGE SENSOR WITH DUAL-ADC SHOT-NOISE-AWARE DYNAMIC RANGE EXTENSION 2117

Fig. 2. VLSI architecture of WDR tunable-color digital pixel sensor prototype.

PMT. In contrast, CCDs can be employed in an arrayed imple-mentation, but do not allow for on-chip integration of peripheralcircuits such as for signal conditioning. This increases cost andlimits miniaturization. CMOS technology, on the other hand,has the advantages of low cost, high integration density, andsignal processing versatility, as for example demonstrated ina time-resolved fluorescence imager [11] and a lab-on-chipfluorometer [12].Numerous dynamic range (DR) enhancement circuit tech-

niques for CMOS image sensors have been reported. Thelogarithmic sensor provides a wide DR with a simple circuitimplementation but achieves a low overall linearity and poorsensitivity under high illumination [13]. The multiple-capturesensor provides a wide DR and maintains high linearity butresults in SNR drops at the high illumination range [14]. Theasynchronous self-reset technique, albeit requiring a largepixel area, extends the DR and simultaneously achieves highlinearity, SNR, and a sensitivity comparable to that of the activepixel sensor [15], [16]. Combined high SNR, DR, and linearityare often primary design requirements for biosensors.Early asynchronous self-reset based prototypes [17] output

only the self-reset count but neglected the charge that remains inthe integration capacitor, commonly referred to as the well. Thisresidue charge is not read out and introduces an error. To miti-gate this shortcoming, residue quantization is introduced to theasynchronous self-reset technique [15], analogous to a two-stepsubranging analog-to-digital converter (ADC). The subrangingADC consists of a self-reset ADC, referred to as a coarse ADC(cADC), which produces the most significant bits (MSBs), fol-lowed by a residue ADC, referred to as a fine ADC (fADC),which produces the least significant bits (LSBs). The method ofresidue quantization in [15] is by reusing the in-pixel ADC toalso process a signal beyond the full well capacity. Since thefADC and cADC have different input ranges, using the same

ADC circuit in the presence of input-dependent shot noise doesnot lead to an overall noise-optimized design.We present a CMOS tunable-color wide dynamic range

image sensor prototyped in a standard 0.35 m CMOS tech-nology. The sensor integrates an 8 8 array of pixels utilizingthe CMOS tunable-color photogate (CPG), with the earlier-gen-eration single-pixel prototypes reported by us in [10], [18],[19]. The CPG employs the polysilicon gate as an optical filter,thus requiring no additional optical color filter. We also presentan analysis that accounts for photon shot noise for the archi-tectural design of a two-step ADC for image sensors. Basedon this analysis, the sensor implements a WDR asynchronousself-reset readout architecture that places the residue ADCat the column level. The sensor is experimentally validatedthrough the measurement of color light intensities and through2D color imaging. It is also integrated into a contact imagingmicrosystem for sensing fluorescent samples in a microfluidicchannel.The rest of the paper is organized as follows. Section II

discusses the overall VLSI architecture. Section III reviewsthe tunable-color photogate device. Section IV quantitativelyanalyzes the two-step ADC VLSI architecture in the presenceof shot-noise. Sections V and VI discuss the circuit imple-mentation of the prototype and reports experimental results,respectively. Section VII describes the experimental validationin fluorescence imaging of samples in a microfluidic channel.Section VIII highlights key observations.

II. VLSI ARCHITECTURE

Fig. 2 depicts the chip-level VLSI architecture of the imager.The pixel is schematically depicted in Fig. 2(b). Each pixelintegrates a 50 m 50 m tunable-color photogate (CPG)[19] for color sensing and a / -body photodiode to providemonochromatic sensing.


The sensor implements the asynchronous self-reset with theresidue readout technique. In this scheme, the photocurrent isfirst estimated by a coarse ADC and the estimation error issubsequently quantized by a fine ADC and used to improvethe accuracy of the final result. In the presented implementa-tion, the asynchronous self-reset cADC is located in the pixelbut the fADC is implemented as a column-parallel single-slopeADC. This enables the decoupling of specifications for the twosub-ADCs for noise-optimization described in Section IV. Theerror that results from coarse analog-to-digital conversion, i.e.,the residue charge, is buffered by an in-pixel source followerbefore it is fed into the column-level fADC.Fig. 3 depicts key signals within the imager. Analog-to-dig-

ital conversion is divided into two phases. Phase 1 is a globaloperation which begins at the start of the integration time, whenthe ‘cADC RESET’ signal is asserted to reset the integrationnode and in-pixel counters. The shutter then closes, feeding thephotocurrent into the readout circuit, and the photodiode outputvoltage raises. If illumination is large enough to exceedthe well capacity, the comparator generates a pulse at the node

, which turns on the reset transistor to reset . Eachreset increments the in-pixel counter by one. The number ofresets corresponds to the MSBs of the overall two-step ADCoutput. At the end of integration time, phase 2 begins. In phase 2,the ‘fADC RESET’ signal is asserted to indicate the start of theresidue digitization using single-slope operation. For each con-version, a voltage ramp is fed into each column-parallelcomparator to be compared against the residue voltage, . As

reaches , the fADC counter latches in the currentvalue of the global counter. The result of this phase producesthe LSBs of the final output. Since the fADC is implemented asa column-parallel ADC, it goes through the sampled residues inthe pixels within a column and digitizes them sequentially. Inother words, light exposure and phase 1 conversion are globaloperations whereas the phase 2 conversion of residues is per-formed sequentially.An on-chip R-2R digital-to-analog converter (DAC) is used

to successively generate multiple control voltages for the CPGs(described in Section III) and generate the voltage ramp for thecolumn-parallel single-slope ADCs (described in Section V).

III. CMOS TUNABLE-COLOR PHOTOGATE

Conventional filterless color sensing techniques that solelyrely on integrated circuit process technology are based onsensing at specific depths in the bulk silicon. As a result,these techniques tend not to scale well with technology [20],[21]. The color (e.g., RGB Bayer mosaic) filter array used incommercial color cameras does not offer the flexibility to tunedetection wavelengths. The CMOS tunable-color photogate(CPG) has been developed to mitigate these difficulties byemploying the polysilicon gate as an optical filter. Thus, theCPG does not require an additional optical color filter. Anearlier generation single-pixel prototype has been reported in[10], [18] with detailed analysis reported in [19]. A concisedescription of the CPG principle of operation is given in thissection.

Fig. 3. Transient of key signals in the imager prototype.

A. Principle of Operation

When the CPG is illuminated, the absorption of light is de-scribed by the Beer–Lambert law [22]. The absorbed photonsgenerate electron-hole pairs, giving rise to a photocurrent for asingle wavelength input that is given by

(1)

where is the radiation intensity, is the elementary charge,is the area of the detector, is the wavelength, is Planck’s

constant, is the speed of light in vacuum, is the absorptioncoefficient, is the effective depth of the sensing region, and

is the light absorption from the various layers between thelight source and the detection volume. is a function of andthe control parameter determines the value of . For a givendetector size, (1) can be rewritten as

(2)

When multiple wavelengths of light are incident simultane-ously, the intensities at these wavelengths can be determined bymeasurements frommultiple photo detectors [23]. For example,for a two-wavelength input, the photo currents and mea-sured by two photo detectors can be related to the input intensi-ties and (at and , respectively) by

(3)

(4)


Fig. 4. Cross-sectional view of the CMOS tunable-color photogate (CPG). Thepoly-gate provides wavelength-dependent light absorption.

where the -coefficients describe the transfer function of thedetectors and can be obtained empirically. Specifically, isthe detector sensitivity at with control input . The input in-tensities and can be obtained by solving the system ofequations, provided that the detectors have unique spectral re-sponses (i.e., (3) and (4) are linearly independent). This modelcan be extended to a finite set of wavelengths. To determinethe intensity of an input spectrum to a resolution of distinctwavelengths, measurements are required from detectors.Provided that the wavelengths are well-separated, this methodoffers the flexibility to tune to any arbitrary set of wavelengthswithin the sensitive range of the silicon photodiode. However,the major limitation of this approach is that it requires the com-plete set of wavelengths to be sensed be known a priori sothat the appropriate -coefficient model can be developed. As acounterexample, the sensor would report incorrect intensities ifthree wavelengths are present at the input but only a two-wave-length model is used for reconstruction.

B. CPG VLSI Implementation

To create the equivalent of multiple photo detectors withunique spectral responses, the CPG depicted in Fig. 4 has beendeveloped, comprising core and edge regions. The core regionof the CPG is covered by a polysilicon gate. A -diffusion,referred to as the edge region, forms the device output. A-diffusion fabricated in an -body forms the body bias

ohmic contact.The gate performs two key functions for color sensing. First,

it functions as an optical filter to provide wavelength-dependentabsorption as described above. Second, it is a terminal for the in-duction of an electric field to modulate the extent of photo-gen-erated carrier collection in the core region, the area under thegate. When is applied such that no depletion region isformed under the gate, photo detection only takes place nearthe / -body depletion region. When another is appliedto form a depletion region at the CPG core, it also participatesin photo detection. But the light experiences wavelength-depen-dent absorption as it travels through the gate. Since the gate pro-vides greater attenuation at shorter wavelengths, the core regionprovides additional long-wavelength (e.g., red) responsivity tothe CPG. Since the edge and core of the CPG have differentspectral properties, when different gate-to-body voltages are ap-plied, an equivalent of multiple detectors with unique spectral

Fig. 5. Micrograph of the 0.35 m standard CMOS 2 mm 2 mm color sensordie with a 175 m 175 m pixels.

Fig. 6. Measured 50 m 50 m CPG photocurrent for monochromatic lightat 620 nm (red), 520 nm (green), and 450 nm (blue).

responses is created, e.g., for two colors, (3) and (4) are imple-mented by a single device.It is worth noting that, in a 0.35 m standard CMOS tech-

nology, the thickness of the polysilicon gate is approximately300 nm [24], leading to an approximate attenuation of 65%for blue light (450 nm), 30% for green light (520 nm), and15% for red light (620 nm), based on the wavelength-dependentpolysilicon absorption coefficients of 3.56, 1.35, and 0.45 mfor blue, green, and red light, respectively. The overall imagechip micrograph is depicted in Fig. 5. The photocurrent of a50 m 50 m CPG is depicted in Fig. 6, where the photocur-rent is measured across for three monochromatic illumi-nations using a semiconductor parameter analyser. To highlightthe relative change in the current, the results are normalized toone at V. For V V, the ratio ofthe currents corresponding to each color changes significantlyacross . Therefore, V and V are uti-lized for the multiple measurements as required in (3) and (4).The generation of multiple values of is automated by theon-chip DAC. Multiple CPG responses at different values of


are passed through the ADC for digitization. The recon-struction algorithm that solves for the input spectrum in (3)and (4) is implemented in software.

IV. SHOT NOISE-AWARE WDR TWO-STEP IMAGER ADC

The exploitation of photon shot noise to reduce the noise re-quirement of a multi-ramp single-slope imager ADC has beendemonstrated [25]. In this section, shot noise exploitation forthe asynchronous self-reset with residue readout WDR archi-tecture is described qualitatively first. A quantitative analysis isthen presented for a general two-step imager ADC.

A. Qualitative Analysis

In Fig. 7, the transfer characteristic of the image sensor isplotted on a logarithmic scale. Under the customary assump-tion that one input photon results in one output electron, theinput-referred and output-referred quantities are equivalent andare interchangeable for the purpose of the following analysis.Noise sources depicted in Fig. 7 are the photon shot noise andread noise, which consists of thermal and flicker noise of thephotodiode and readout circuit. In a conventional active pixelsensor (APS), the sensor output increases linearly with the lightintensity until the full well capacity is reached. For animager with extended DR, the output reaches a higher value,

, typically limited by mechanisms that are specific to im-plementation, e.g., in-pixel memory depth for certain self-resetschemes.The photon shot noise has a standard deviation equal to the

square root of the input light signal, in units of electrons. Be-cause photon shot noise increases with the input, as opposed tothe input-independent read noise, it becomes the dominant noisesource at higher light intensities. In this part of the input range,the conventional ADC has a better noise performance than is re-quired, i.e., its quantization and thermal noise can be increasedwithout decreasing the overall noise performance.Fig. 8 also illustrates the above idea for the asynchronous self-

reset cADC and residue fADC architecture depicted in Fig. 2. Inthis architecture, the cADC is only active when the input signalexceeds the full well capacity, , at which point, the shotnoise is . Therefore, since the irreducible shot noise com-ponent is already substantial, a cADC noise floor much below

is over designed in terms of keeping the combined shotnoise and read noise at a reasonable level. Unlike the cADC,the fADC operates from the dark condition to . Therefore,the fADC noise floor is ideally minimized. Since the area con-straint limits the performance of in-pixel ADCs and that a highperformance is required from the fADC, it is implemented inthe periphery as a column-level circuit.

B. Quantitative Analysis

This section presents an analysis that can be used as a guide-line for the architectural design of the two-step ADC. In orderto keep the analysis general to any two-step imager ADC archi-tecture, the formulation is not specific to particular fADC andcADC implementations. Therefore, it is assumed that the noisefloors of the ADCs are independent of the input amplitude, asmost types of ADC exhibits this characteristic.

Fig. 7. Imager transfer curve illustrating the concept of shot noise as an input-dependent noise source.

Fig. 8. Imager transfer curve illustrating shot-noise-aware design.

The analysis begins with the well capacity, which is theamount of charge that a pixel can hold in a single photocurrentintegration, given by

(5)

where and are the integration capacitance and inte-gration voltage, respectively, and is the elementary charge. Itis worth noting that is in most cases the parasitic capac-itance of the photodiode. Since is process dependent and

depends primarily on the supply voltage, is largelydetermined by the process technology.The asynchronous self-reset with residue readout scheme can

be regarded as a conventional active pixel sensor with dynamicrange extension at high illumination via self-reset operation.Since imager ADCs are usually designed such that their quanti-zation noise does not exceed the read noise [25], in this analysis,the fADC root-mean-square (RMS) quantization noise, , ischosen to be equal to the read noise, , i.e.,

(6)


It is well established that, for an input that follows a uniformprobability distribution function, the RMS quantization noise ofan ADC, , can be related to the quantization step size, ,as follows [26]

(7)

Therefore, the fADC quantization step size is given by

(8)

From the ratio of the largest to the smallest signal, the reso-lution of the fADC and cADC can be readily formulated. Thenumber of bits for the fADC, , is given by

(9)

The saturation signal of the entire two-step DR, , is deter-mined by the maximum input light intensity specification. Thenumber of bits for the cADC, , is given by

(10)

The resolution of the two-step ADC is therefore bits.The analysis next proceeds to obtain an expression for the

RMS quantization noise of the cADC, . The standard devi-ation of shot noise is given by

(11)

where is the input photon count. Referring to Fig. 8, inorder to constraint the cADC quantization noise to an accept-able level, , a quality ratio is defined to relate thecADC noise to shot noise, given by

(12)

If , then . Therefore, to ensure thatis below shot noise, is chosen to be less than unity.

As depicted in Fig. 8, since has its minimum atwithin the range that the cADC self-resets (i.e., to) and since must be smaller or equal to ,is chosen to be equal to the acceptable noise level evaluated

at the full well level, i.e.,

(13)

This guarantees that is below shot noise for .At full well, using (11) and (12), is given by

(14)

Using (7) and (14), the quantization step size for the cADC isgiven by

(15)

As a measure of the noise increase, or equivalently the reductionof the noise requirement of the cADC, the factor is defined,given by

(16)

which is interestingly proportional to the square root of the wellcapacity. Also, combining (9) and (16), can be expressed as

(17)

which states that once the fine ADC resolution, , is fixed, thefactor is inversely proportional to .Key performance metrics of the fADC and cADC can then

be computed. The dynamic range is defined as the ratio of thelargest signal to the smallest detectable signal. The DRs, in dB,for the fADC and cADC are, respectively

(18)

and

(19)

But as shown in Fig. 8, the DR of the overall two-step ADC isnot merely as is the case for a conventional two-step ADC.Rather, the extended DR is involved, given by

(20)

The DR of the overall two-step ADC is given by

(21)

This DR analysis reveals an important property of the two-step WDR ADC. When shot noise is considered, as in this anal-ysis, the cADC noise floor can be raised, which relaxes the DRrequirement of the cADC. However, as illustrated in Fig. 8,still has to be designed to a level below the cADC LSB stepsize, namely, the full well level. Therefore, the cADC effectivenumber of bits (ENOB) exceeds its actual number of bits. Asmentioned previously, if is designed to be in the neighbor-hood of , the cADC has better noise performance than isrequired. But, if is equal to the full well level, then the finalresult of the entire two-step ADC is only accurate to the cADCLSB and renders the entire bits of the fADC inaccurate.The signal-to-noise ratios (SNRs) in dB for the fADC in the

read noise dominant (low signal) and shot noise dominant (highsignal) regimes are, respectively

(22)


and

(23)

The peak SNR in dB of the fADC is obtained by evaluatingat the highest fADC input, given by

(24)assuming, as it is typical of imagers, that shot noise dominatesat the full well level. Analogously, since is designed (from(13)) to be below shot noise in the extended dynamic range, theimager is shot noise dominant within this range. Hence the SNRin dB in the extended DR is given by

(25)

which has the same expression as . Analogously, thecADC peak SNR in dB is given by

(26)Lastly, the peak SNR of the overall two-step ADC is given by

(27)which is different from the ratio of the maximum input to thenoise at the zero signal level, as is often the case for most typesof ADCs. This difference is a direct result of the presence of shotnoise. Strictly speaking, the above peak SNR expressions areapproximations, as read noise has not been included. But sincethe peak SNRs are evaluated at the illumination level whereshot noise dominates over read noise, the approximations arenonetheless accurate.As a numerical example of the above quantitative analysis,

the following parameters with values typical to CMOS imagersare assumed [27]: , 10,000 , and

. Assume that the fADC quantization noiseis designed to be equal to the read noise, i.e.,

. From (8), . From (9) and (10), thenumber of bits for the fADC and cADC are bits and

bits, respectively. Suppose the cADC is designed tohave a quantization noise comparable to shot noise at the fullwell input, i.e., . Therefore, based on (14) and (15), itcan be computed that and ,respectively. From (16), the cADC noise floor can be raised 20times ( ) while not incurring a severe noise degradation.This translates to a power and/or area saving. From (24) and(26), the peak SNRs for the fADC and cADC are estimated tobe 40 dB and 60 dB, respectively. From (27), the peak SNR ofthe two-step ADC is 60 dB. From (18)–(20), , , and

are 66 dB, 80 dB, and 40 dB, respectively. Since theoverall DR is , the fact that shot noiseis taken into account relaxed the cADC noise requirement by26 dB. An important benefit of this is a much reduced pixelarea for an in-pixel cADC implementation (an asynchronousself-reset ADC is required to be implemented in-pixel). It is

also worth noting that since the shot noise magnitude is equalto the square root of the signal magnitude, in this example, after10,000 resets, the equivalent signal is 10,000 . This resultsin a shot noise of 10,000 , which is equal to the full wellcapacity. Therefore, the entire residue or the entire fADC outputconsists of noise.

V. CIRCUIT IMPLEMENTATION

The design of Fig. 2 has been fabricated in a 0.35 m standarddigital CMOS technology. The 175 m 175 m pixels with a10% fill factor are tiled to form a 8 8 array for imaging.

A. Pixel Circuit Implementation

The pixel readout circuit consists of an electronic shutter, areset block, and ADC block that implements asynchronous self-reset, the coarse part of the two-step ADC operation. The resetblock includes a comparator and a reset transistor. The ADChas a 15-bit linear feedback shift register (LFSR) counter whichdominates the pixel area. The choice of 15-bit is to provide awide dynamic range even without the use of the fADC. Thisenables the imager to support a high frame rate mode with pixel-parallel A/D conversion.Referring to the circuit in Fig. 2(c) and timing diagram

in Fig. 3, in the beginning of each integration period, thecounter is cleared and the photodiode output is chargedto the reset voltage V. The -body is biased atthe voltage V. The photocurrent causesto rise, charging the integration capacitor . Whenreaches the comparator reference voltage V, thecomparator changes state, causing the reset transistorto turn on, resetting to . After reset, the comparatoroutput toggles back to the original state. The combination oftwo toggles generate a pulse with a width that equals the timeit takes to reset . In the absence of circuit nonidealitiesand ignoring the residue, the resulting pulse train has a pulsefrequency proportional to the incident light intensity. Theasynchronous self-reset ADC is less sensitive to supply voltagescaling as it effectively represents the light signal by a digitalcount, rather than a voltage across a capacitor.The voltage comparator is a two-stage design with large

PMOS input transistors to lower thermal and flicker noise.Fig. 9 depicts the voltage comparator. The first stage employscross-coupling to increase the output resistance of the loadtransistors through . The second stage provides anadditional gain. The first stage and the second stage consume18 A and 10 A for the 3.3 V supply, respectively. The com-parator has a 66 dB simulated DC gain for resolving 1 mVfor a maximum input swing of 2 V. It has a 3-dB bandwidth of10 MHz.

B. Column-Parallel Analog-to-Digital Converters

The column-parallel single-slope ADC of Fig. 2(d) digitizesthe residue voltage . It consists of a voltage comparatorwith the same topology as the in-pixel comparator of Fig. 9 anda 15-bit binary counter. A global counter is connected to boththe on-chip DAC and the column-parallel ADC sub-circuits asshown in Fig. 2(a). During analog-to-digital conversion, it incre-ments in order to have the DAC output a ramp voltage


Fig. 9. Schematic of the voltage comparator.

Fig. 10. Schematic of the on-chip R-2R DAC.

Fig. 11. Schematic of the operational amplifier within the DAC.

to the column-parallel ADCs. As depicted in Fig. 3, iscompared to the residue voltage . When reaches

, the comparator clocks the register to latch in the presentglobal counter value, which is the digital representation of .

C. Digital-to-Analog Converter

The schematic of the on-chip DAC is shown in Fig. 10. TheDAC is based on the R-2R architecture [26], due to the avail-ability of high-precision resistors in the CMOS process used.Unlike the conventional resistance-ratio ladder converter [26],the R-2R converter realizes binary-weighted currents with asmaller number of components and with a resistance ratio ofonly two, independent of the number of bits.Switches are sized proportionately to accommodate the bi-

nary increase in the current level through each branch. The unitresistors are 5 k non-silicided polysilicon resistors, eachoccupying an area of 24 m 1.1 m. The opamp is based onthe two-stage opamp architecture, depicted in Fig. 11. PMOS

Fig. 12. Experimentally measured imager output as a function of inputintensity.

input transistors are used to lower the flicker noise. The DACoccupies an area of 300 m 120 m.

VI. EXPERIMENTAL RESULTS

A. Pixel Readout Circuit

The experimentally measured transfer characteristic is de-picted in Fig. 12, obtained from illumination provided by ahalogen lamp onto a pixel. The input light intensity is measuredby an optical power meter (with a detector calibrated for broad-band sensing) and is varied by over four orders of magnitudeusing neutral density (ND) filters. ND filters used are from Thor-labs with optical densities (OD) of 0.3, 0.7, 1.0, 2.0, 3.0, and4.0 and are combined to provided various degrees of attenua-tion. Measurements are collected using the 50 m 50 m CPGwhich is set to the highest photo sensitivity, i.e., V.At each intensity, 32 measurements are obtained to calculatethe average and standard deviation of the imager output. Theexperimentally measured transfer characteristic is depicted inFig. 12, obtained from illumination provided by a halogen lamponto a pixel. The input light intensity is measured by an opticalpower meter (with a detector calibrated for broadband sensing)and is varied by over four orders of magnitude using neutraldensity (ND) filters. ND filters used are from Thorlabs with op-tical densities (OD) of 0.3, 0.7, 1.0, 2.0, 3.0, and 4.0 and arecombined to provided various degrees of attenuation. Measure-ments are collected using the 50 m 50 m CPG which is setto the highest photo sensitivity, i.e., V. At each inten-sity, 32 measurements are obtained to calculate the average andstandard deviation of the imager output.The experimentally measured SNR of the imager output, de-

picted in Fig. 13, is calculated as mean over standard deviationof the output. The input-referred dynamic range is defined as themaximum output over the RMS value of the readout noise ,i.e., the standard deviation of the imager output under dark con-dition [27]. Therefore, the DR is the range between SNRand the highest signal level in Fig. 13 and is measured to be82 dB, limited by the maximum light intensity achievable by


Fig. 13. Experimentally measured imager SNR as a function of input intensity.

Fig. 14. Histogram of the sensor output under uniform illumination.

the light source in the high end and the read noise in the lowend. In Fig. 13, as the input signal increases, the SNR improvesat 20 dB/dec at low illumination where the total noise is domi-nated by the input-independent noise of the fADC readout cir-cuit. As the number of reset operations increases, the noise inthe self-reset loop (e.g., reset noise) accumulates, which reducesthe rate of increase of the SNR to approximately 6 dB/dec atvery high illumination levels. This increase in cADC noise hasalso been reported in [15] and must be minimized. As depictedin Fig. 13, the peak SNR of 46.2 dB has been measured at thehighest input intensity. Fig. 13 also depicts a measured numberof reset pulses of 200, which is equivalent to a DR increase of46 dB over the same CMOS image sensor without any DR en-hancement, i.e., with the cADC disabled.To evaluate the fixed pattern noise (FPN) of the imager, uni-

form illumination is applied to the entire pixel array. Fig. 14depicts the histogram of image intensities resulting from uni-form illumination. The average intensity is 251.1 counts with aFPN ( ) of 0.38%.

B. Digital-to-Analog Converter

The DAC consumes 1.24 mW and achieves 8-bit accuracywith INL and DNL shown in Fig. 15(a) and (b), respectively.

Fig. 15. Experimentally measured on-chip DAC performance: (a) INL, and(b) DNL.

The opamp within the DAC achieves a simulated DC gain of69 dB and a 3-dB bandwidth of 20 kHz.

C. System-Level Validation in Color Light Measurements

The 0.35 m CMOS prototype in Fig. 5 has been tested inlight intensity measurements at the green (520 nm) and red (620nm) wavelengths using two current-controlled light-emittingdiodes (LEDs) for input illumination.In order to measure the intensity at two known wavelengths,

according to (3) and (4), an empirical model with four -coef-ficients is required. The extraction of -coefficients can be per-formed as follows. For example, in (3), to extract , a knownlight intensity serves as the input of the measurement (at

). Similarly, for , a known light intensity is appliedas an input for another measurement at . This process isthen repeated for . Following the above procedure, onlymeasurements are required to determine all -coefficients.

Additionally, it has been found that modeling accuracy can beimproved by simultaneously utilizing multiple combinations ofinput colored light intensities to solve for the average -coeffi-cients. The -coefficients are obtained only once, and are storedfor subsequent reconstruction calculations.To resolve the input to two wavelengths, each input light is

measured two times using V and V. Theraw measurements and the previously obtained model are com-bined to reconstruct the input using (3) and (4). Fig. 16 depictsmeasured intensities after reconstruction for an illumination thatsimultaneously contains green (520 nm) and red (620 nm) light.For each of the two wavelengths, intensities of 0, 60, 120, 180,240, and 300 W cm have been used. Fig. 16(a) depicts mea-sured tunable-color photogate response across the illuminationrange for two gate-to-body voltages. These data are used to de-termine the -coefficients to create a linear model (depicted as amesh). In order to evaluate the crosstalk between color channels,for each intensity step, the intensity of the other color is swept.For example, in Fig. 16(b), for each of the six green intensities,the red intensity is varied in six levels. Therefore, based on six


Fig. 16. Simultaneous two-color light-emitting diode (LED) illumination measurements: (a) measured tunable-color photogate response for two gate-to-bodyvoltages across illumination (circles). A linear model based on -coefficients is superimposed (mesh). (b) Reconstructed intensity of the green (520 nm) inputcomponent. (c) Reconstructed intensity of the red (620 nm) input component.

Fig. 17. Color imaging: RGB image reconstructed from input images captured at V, 0.3 V, and 0.6 V, from top to bottom, respectively.V.

intensity steps, each against six intensities of the other color, atotal of 36 measurements have been performed. Each error barcontains a sweep across all intensities of the other color and de-picts one standard deviation from the mean value. Fig. 16(c) isanalogous to Fig. 16(b) but for the red component of the inputlight.The SNR is commonly defined for imagers as ,

where and are the mean and standard deviation of the outputcalculated over temporal measurements for all pixels [28]. Asshown in Fig. 16(b) and (c), the peak SNRs measured at the in-tensity of 300 W cm are 24.3 dB and 28.5 dB for the greenand red components, respectively. In a separate high input in-tensity test (result not shown), the imager achieves a peak SNRof 29.2 dB and 34.8 dB for green and red light at the intensityof 1500 W cm (limited by maximum LED output).

D. System-Level Validation in 2D Color Imaging

Although the imager has been designed for fluorescencesensing application rather than for the photographic applica-tion, its ability to reproduce an image is evaluated via capturingof a still photographic image. It is an image abbreviating the

string ‘University of Toronto’ by the characters ‘U’, ‘o’, and‘T’, on a black background. The approximately 10 cm 10 cmimage is held approximately 0.9 m from the lens, which fo-cuses the light onto the pixel array. The lens aperture is atF/16 and the integration time is 1 sec under W cmof illumination. The image captured by the proposed sensoris depicted in Fig. 17. The entire array of pixels use a globalcolor model or -coefficient matrix. One key advantage of theproposed approach to color sensing is that each pixel producesthe entire set of RGB values. Therefore, color interpolation,a process that approximates missing color information fromneighboring pixels commonly performed in cameras with acolor filter mosaic array, is not required. This eliminates theassociated color artifacts.Table I summarizes the experimentally measured electrical

characteristics of the image sensor prototype depicted in Fig. 5.

VII. VALIDATION IN FLUORESCENCE IMAGING

The tunable-color sensor prototype has been evaluated as apart of a fluorescence imaging microsystem to validate its suit-ability for point-of-care (POC) diagnostic applications. POC


TABLE IEXPERIMENTAL CHARACTERISTICS

devices are becoming increasing popular as they promise tobring diagnostic technology from the standard laboratory set-ting to the patient residence to facilitate early diagnosis [29].Although miniaturization is a key for the development of suchdevices, for optical transduction such as fluorescence-based de-tection, a fluorescence microscope is commonly employed. De-spite the high sensitivity and selectivity offered by this trans-duction method and its widespread applications ranging fromthe detection of nucleic acids, proteins and small molecules, itsincorporation into POC devices has been limited due to the lim-ited portability and the high cost of the instrumentation. Oneemerging technique with a potential to overcome the limitationsof a fluorescence microscope is contact imaging [30]. Unlikethe conventional fluorescence microscope, in contact imagingas depicted in Fig. 1, the object to be imaged is placed in closeproximity to the focal plane, eliminating the need for bulky andexpensive optics such as a system of lenses and mirrors, whichenables miniaturization to realize lab-on-a-chip platforms.Microfluidic networks offer many advantages for chemical

and biological sensing. First, reaction time is greatly shortened,in some cases from hours to minutes [31], as active deliveryby electrokinetic flow can be used to accelerate interactionsbetween molecules over an otherwise slow diffusion-limitedprocess. Secondly, small sample volumes in the nano-liter rangecan be readily transported and processed by means of microflu-idic networks. Thirdly, sensing of samples within a microfluidicchannel where the chemical reaction occurs facilitates real-timedetection.

Integrating an imager with microfluidics can serve a varietyof applications. Spatial imaging of a fluidic channel, for ex-ample, can be a method to analyze the result of electrophoresisexperiments where the outcome is determined by detecting thedistance traveled by dispersed particles relative to a fluid underthe influence of an electric field [32].Quantum dots (QDs) as fluorescent markers exhibit a number

of unique optical properties that render them superior than or-ganic fluorescent dyes. These unique properties include: narrow,symmetric and size-tunable emission spectra (full width at halfmaximum, FWHM of 25–35 nm); strong and broad absorptionspectra; high quantum yield ( 20%) and long life time ( 10 ns)[8]. As compared to organic fluorophores, QDs have greater re-sistance to photobleaching that enables long-term monitoring.The broad absorption spectra of QDs allows for multiple colorsof QDs to be excited efficiently with a single excitation sourcewhich is generally not possible with organic dyes. These prop-erties make QDs ideal as fluorescent biomarkers.

A. Microsystem Prototype Design

The microsystem prototype consists of a blue LED forfluorescence excitation, an optical filter for excitation rejection,a fluidic structure for holding the sample solution, and theCMOS CPG imager for photo detection. A 100- m-thick,1.5 mm 1.5 mm optical interference filter is used to attenuatethe 450 nm ( nm) excitation light from the PhilipsLuxeon K2 450 nm ( nm) blue LED. The filteris fabricated using 60 layers of Nb O and SiO (by OmegaOptical) to the required specification, and optically tested priorto integration with the CMOS die. This approach is chosenover the direct deposition of thin-film layers over the CMOSdie to ensure that well-established methods for coating planarsubstrates can be used during filter fabrication. The filter is along-pass design with a cut-off wavelength of 510 2 nm. Thefilter has been tested to provide an optical density (OD) of six(i.e., 10 attenuation) at the excitation wavelength of 450 nm,with a transmission rate greater than 90% at 520 nm and onaverage greater than 85% from 520 nm to 700 nm.The microfluidic device consists of a hybrid of top poly-

dimethylsiloxane (PDMS) cover and bottom glass substrate.The channels are fabricated in PDMS using a soft-lithography(rapid prototyping and replica molding) technique. PDMS baseand curing agent are thoroughly mixed in a 10:1 ratio, themixture has been degassed under vacuum, and then 3 g of themixture are poured onto the microfluidic template and cured inan oven at 120 C for 30 min. The cured PDMS cover is peeledoff and the inlets and outlets at the ends of each channel arepunched out using a 2 mm diameter metal bore. The PDMScover is then air plasma oxidized for 30 s at 10.5 W and isimmediately sealed to a plasma oxidized glass coverslip.Fig. 1 depicts a simplified cross-sectional view of the re-

sulting microfluidic device. It has 1 cm (length) 250 m(width) 11 m (height) channels, terminated by an inlet andan outlet on each end. The microfluidic device is subsequentlyintegrated with the CMOS sensor as shown in Fig. 18(a).Fig. 18(b) depicts an enlarged view of the microfluidic channelpassing over the sensor pixel array.


TABLE IICMOS FLUORESCENCE MICROSYSTEMS COMPARATIVE ANALYSIS

Fig. 18. Photograph of the microfluidic device placed over the CMOS sensordie: (a) overall configuration showing entire package cavity, and (b) close-upview of fluidic channel running across the CMOS die.

B. Fluorescence Contact Imaging Experimental Results

To evaluate the applicability to fluorescence imaging, theproposed microsystem is utilized to image QDs in a microflu-idic device. Fig. 19 show measured results from fluorescenceimaging as captured by the presented CMOS image sensor.Fig. 19(a) captures the fluorescence of 2 m red QDs (peakemission wavelength at 620 nm) in a microfluidic channel.Since only a single color of emission is to be sensed, the CPGfunctions as a monochromatic detector (i.e., reconstruction isnot necessary) and is set to V to maximize sensitivity.The samples have been imaged under an excitation power ofapproximately 0.5 mW mm and an exposure time of 10 sec.The presence of the fluidic network introduces light scattering,which combined with the stray LED output in the filter pass-band, resulted in a background signal of approximately 250sensor output codes. To remove the background signal com-ponent, the background signal is subtracted from the originalimage to produce the result in Fig. 19(a). Fig. 19(b) depictsa background subtracted image of a spot (diameter 1 mm)of 2 m solution of red QDs. The sample solution is directlyspotted on the thin-film filter, and as a result, the sample spotis approximately 100 m away from the detector due to thethickness of the filter. To highlight the features in the capturedimages, Fig. 19(a) and (b) are intensity-thresholded to produceFig. 19(c) and (d), respectively. Fig. 19(c) highlights the fact

Fig. 19. Fluorescence imaging of 2 m red quantum dot in solution phase:(a) background subtracted image of sample in a 250 m-wide microfluidicchannel, and (b) background subtracted image of 1 mm-diameter rQD spotdeposited directly on the thin-film filter. (c) and (d) are intensity-thresholdedimages of (a) and (b), respectively.

that the fluorescence intensity is higher on the left hand sideof the image. This is due to a concentration gradient of thequantum dot solution, which has been injected into the channelfrom the inlet located on the left of the image. Thresholdinghas been performed in software but can be readily implementedon-chip [5]. Table II compares the proposed work to recentlyreported CMOS fluorescence imaging microsystems.It is often meaningful to characterize the detection limit by

the required sample size, rather than solely by the analyte con-centration [35]. Since 25 nL of sample volume has been injectedinto the channel with a 2 m concentration, the microsystem isable to detect 50 fmol of rQD fluorophore. This is advantageousas fewer time-consuming polymerase chain reaction (PCR) cy-cles are needed to bring the concentration of the target DNA toa level that can be detected.

VIII. DISCUSSION

Spatial resolution is often traded off for SNR and DR in sci-entific imagers. For example, in a DNA detection biosensor,


the pixel is designed to be pitch-matched to an individual spoton a microarray [1]. The presented pixel has a large area of175 m 175 m. The rather low fill-factor of 10% is due to theinclusion of a digital counter for asynchronous self-reset oper-ation. To improve fill factor, one approach is to implement thecomparator and counter on the column level to perform syn-chronous reset operation, at the expense of lowered SNR due todelayed reset [36].The requirement that the input spectra be known a priori as

discussed in Section III-A is not a serious limitation when thesensor is applied to fluorescence sensing as the spectrum typi-cally consists of only the emission wavelengths and these wave-lengths can be made well-separated by an appropriate choice offluorophores. The tunability of the sensor is a key advantage asthis allows the detection of specific emission wavelengths fromvarious fluorophores.The CMOS die is housed inside a 400- m-deep cavity within

an integrated circuit package. The microfluidic network is sus-pended on top of the cavity, as depicted in Fig. 18(a). The totaldistance from the sample to the detector, including the thicknessof the reservoir bottom, is approximately 300 m. This distanceleads to a reduced photon collection efficiency compared to de-positing the fluorophores on the thin-film filter, which is placeddirectly on the CMOS die surface. This is the reason for thehigher intensity in Fig. 19(b) as compared to Fig. 19(a). Thedistance also leads to blurring of the images. Aside from intro-ducing optics such as microlenses to focus the image onto thephoto detectors, a possible solution is to design the microfluidicdevice with an additional lower layer that can be extended downinto the chip package cavity [37]. Thus, channels can be routedto this layer to bring the sample solution closer to the pixel array.

IX. CONCLUSION

A wide dynamic range CMOS tunable-color image sensor ispresented. The sensor integrates an 8 8 array of tunable-colorphotogates. It exploits the wavelength-dependent optical ab-sorption properties of the polysilicon gate to yield color discrim-ination on a standard digital CMOS process without an externalcolor filter array. An analysis is presented for the asynchronousself-reset with residue readout ADC architecture where photonshot noise is taken into consideration. An implementation ofthis architecture is described where the coarse asynchronousself-reset operation and fine residue quantization are performedwith separate circuits, on and off the array, respectively, to yielda noise-optimized design. A prototype is fabricated in a stan-dard 0.35 m CMOS process and is validated in color lightmeasurements. Contact imaging of quantum dot nanoparticleswithin a microfluidic channel validates the prototype in fluo-rescence-based analyte detection. The prototype demonstratestechnologies that enable miniaturized, low-cost bio-sensing formedical diagnostics applications.

REFERENCES[1] H. Eltoukhy, K. Salama, and A. Gamal, “A 0.18- m CMOS biolumi-

nescence detection lab-on-chip,” IEEE J. Solid-State Circuits, vol. 41,no. 3, pp. 651–662, Mar. 2006.

[2] R. Mazurczyka, J. Vieillarda, A. Bouchardb, B. Hannesa, and S.Krawczyka, “A novel concept of the integrated fluorescence detectionsystem and its application in a lab-on-a-chip microdevice,” Sens.Actuat. B: Chem., vol. 118, no. 1-2, pp. 11–19, Oct. 2006.

[3] X. Gao, Y. Cui, R. Levenson, L. Chung, and S. Nie, “In vivo cancertargeting and imaging with semiconductor quantum dots,” NatureBiotechnol., vol. 22, no. 8, pp. 970–976, 2004.

[4] O. Tigli, L. Bivona, P. Berg, andM. Zaghloul, “Fabrication and charac-terization of a surface-acoustic-wave biosensor in CMOS technologyfor cancer biomarker detection,” IEEE Trans. Biomed. Circuits Syst.,vol. 4, no. 1, pp. 62–73, Feb. 2010.

[5] R. Singh, D. Ho, A. Nilchi, G. Gulak, P. Yau, and R. Genov, “A CMOS/thin-film fluorescence contact imaging microsystem for DNA anal-ysis,” IEEE Trans. Circuits Syst. I, vol. 57, no. 5, pp. 1029–1038, May2010.

[6] W. Algar and U. Krull, “Developing mixed films of immobilizedoligonucleotides and quantum dots for the multiplexed detectionof nucleic acid hybridization using a combination of fluorescenceresonance energy transfer and direct excitation of fluorescence,”Langmuir, vol. 26, no. 8, pp. 6041–6047, Apr. 2010.

[7] E. Graves, R. Weissleder, and V. Ntziachristos, “Fluorescence molec-ular imaging of small animal tumor models,” Current Molecular Med.,vol. 4, no. 4, pp. 419–430, Jun. 2004.

[8] M. Bruchez, M. Moronne, P. Gin, S. Weiss, and A. Alivisatos, “Semi-conductor nanocrystals as fluorescent biological labels,” Sci., vol. 281,no. 5385, pp. 2013–2016, 1998.

[9] A. Tavares, M. Noor, C. Vannoy, W. Algar, and U. Krull, “On-chiptransduction of nucleic acid hybridization using spatial profiles of im-mobilized quantum dots and fluorescence resonance energy transfer,”Anal. Chem., vol. 84, no. 1, pp. 312–319, Jan. 2012.

[10] D. Ho,M. Noor, U. Krull, G. Gulak, and R. Genov, “Single-filter multi-color CMOS fluorescent contact sensing microsystem,” in Proc. IEEEInt. Symp. Circuits and Systems (ISCAS), May 2012.

[11] T. Huang, S. Sorgenfrei, P. Gong, R. Levicky, and K. Shepard, “A0.18- m CMOS array sensor for integrated time-resolved fluorescencedetection,” IEEE J. Solid-State Circuits, vol. 44, no. 5, pp. 1644–1654,May 2009.

[12] N. Nelson, D. Sander, M. Dandin, S. Prakash, A. Sarje, and P. Abshire,“Handheld fluorometers for lab-on-a-chip applications,” IEEE Trans.Biolog. Circuits Syst., vol. 3, no. 2, pp. 97–107, Apr. 2009.

[13] S. Kavadias, B. Dierickx, D. Scheffer, A. Alaerts, D. Uwaerts, andJ. Bogaerts, “A logarithmic response CMOS image sensor withon-chip calibration,” IEEE J. Solid-State Circuits, vol. 35, no. 8, pp.1146–1152, Aug. 2000.

[14] X. Liu and A. E. Gamal, “Synthesis of high dynamic range motionblur free image from multiple captures,” IEEE Trans. Circuits Syst. I:Fundam. Theory Applicat., vol. 50, no. 4, pp. 530–539, Apr. 2003.

[15] D. Park, J. Rhee, and Y. Joo, “A wide dynamic-range CMOS imagesensor using self-reset technique,” IEEE Electron Device Lett., vol. 28,no. 10, pp. 890–892, Oct. 2007.

[16] A. Belouchrani and K. Abed-Meraim, “Wide-dynamic-range CMOSimage sensors-comparative performance analysis,” IEEE Trans. Elec-tron Devices, vol. 56, no. 11, pp. 2446–2461, Nov. 2009.

[17] A. Bermak, “A CMOS imager with PFM/PWM based analog-to-dig-ital converter,” in IEEE Int. Symp. Circuits and Systems (ISCAS), May2002, pp. 53–56.

[18] D. Ho, G. Gulak, and R. Genov, “CMOS field-modulated color sensor,”in Proc. IEEE Custom Integrated Circuits Conf. (CICC), Sep. 2011.

[19] D. Ho, M. Noor, U. Krull, G. Gulak, and R. Genov, “CMOS tun-able-wavelength multi-color photogate sensor,” IEEE Trans. Biomed.Circuits Syst., accepted for publication.

[20] X. Fang, V. Hsiao, V. Chodavarapu, A. Titus, and A. Cartwright,“Colorimetric porous photonic bandgap sensors with integratedCMOS color detectors,” IEEE Sensors J., vol. 6, no. 3, pp. 661–667,Jun. 2006.

[21] H. Ishii, Y.Maruyama, H. Takao, M. Ishida, and K. Sawada, “Improve-ment in filter-less fluorescence sensor capability by optimization of po-tential distribution,” in Proc. 4th Asia-Pacific Conf. Transducers andMicro-Nano Tech., Jun. 2008, pp. 68–71.

[22] S. Dimitrijev, Principles of Semiconductor Devices, 1st ed. Oxford,U.K.: Oxford Univ. Press, 2006.

[23] A. Belouchrani, K. Abed-Meraim, J. Cardoso, and E. Moulines, “Ablind source separation technique using second-order statistics,” IEEETrans. Signal Process., vol. 45, no. 2, pp. 434–444, Feb. 1997.

[24] H. Wong, “Technology and device scaling considerations for CMOSimagers,” IEEE Trans. Electron Devices, vol. 43, no. 12, pp.2131–2142, Dec. 1996.

[25] M. Snoeij, A. Theuwissen, K. Makinwa, and J. Huijsing, “Multiple-ramp column-parallel ADC architectures for CMOS image sensors,”IEEE J. Solid-State Circuits, vol. 42, no. 12, pp. 2968–2977, Dec. 2007.

[26] D. Johns and K. Martin, Analog Integrated Circuit Design, 1st ed.New York: Wiley, 1997.


[27] J. Nakamura, Image Sensors and Signal Processing for Digital StillCameras, 1st ed. Boca Raton, FL, USA: CRC Press, 2006.

[28] K. Murari, R. Cummings, N. Thakor, and G. Gauwenberghs, “ACMOS in-pixel CTIA high-sensitivity fluorescence imager,” IEEETrans. Biomed. Circuits Syst., vol. 5, no. 5, pp. 449–458, Oct. 2011.

[29] J. Choi, G. Beaucage, J. Nevin, J. Lee, A. Puntambekar, and J. Lee,“Disposable smart lab on a chip for point-of-care clinical diagnostics,”Proc. IEEE, vol. 92, no. 1, pp. 154–173, Jan. 2004.

[30] J. Honghao, D. Sander, A. Haas, and P. Abshire, “Contact imaging:Simulation and experiment,” IEEE Trans. Circuits Syst. I, vol. 54, no.8, pp. 1698–1710, Aug. 2007.

[31] M. Noor and U. Krull, “Microfluidics for the deposition of density gra-dients of immobilized oligonucleotide probes; developing surfaces thatoffer spatial control of the stringency of DNA hybridization,” Anal.Chim. Acta, vol. 708, no. 1-2, pp. 1–10, Dec. 2011.

[32] M. Behnam, G. Kaigala, M. Khorasani, S. Martel, D. Elliott, and C.Backhouse, “Integrated circuit-based instrumentation for microchipcapillary electrophoresis,” IET Nanobiotechnol., vol. 4, no. 3, pp.91–101, Sep. 2010.

[33] Y. Maruyama, K. Sawada, H. Takao, and M. Ishida, “A novel filterlessfluorescence detection sensor for DNA analysis,” IEEE Trans. ElectronDevices, vol. 53, no. 3, pp. 553–558, Mar. 2006.

[34] M. Beiderman, T. Tam, A. Fish, G. Jullien, and O. Yadid-Pecht, “Alow-light CMOS contact imager with an emission filter for biosensingapplications,” IEEE Trans. Biomed. Circuits Syst., vol. 2, no. 3, pp.193–203, Sep. 2008.

[35] K. Peck, L. Stryer, A. Glazer, and R. Mathies, “Single-molecule flu-orescence detection: Autocorrelation criterion and experimental real-ization with phycoerythrin,” Proc. Nat. Acad. Sci., vol. 86, no. 1, pp.4087–4091, Feb. 1989.

[36] S. Kavusi and A. E. Gamal, “Quantitative study of high dynamic rangeimage sensor architectures,” IEEE Trans. Electron Devices, vol. 43, no.12, pp. 2131–2142, Dec. 1996.

[37] M. Nazari, H. Jafari, L. Leng, A. Guenther, and R. Genov,“192-channel CMOS neurochemical microarray,” in Proc. IEEECustom Integrated Circuits Conf. (CICC), Sep. 2010.

Derek Ho (S’09) received the B.A.Sc. (first class)and M.S. degrees from the University of BritishColumbia, Vancouver, BC, Canada, in 2005 and2007, respectively. He is currently pursuing thePh.D. degree in the Department of Electrical andComputer Engineering, University of Toronto,Toronto, ON, Canada.His research interests are in mixed-signal VLSI

circuits and systems for sensory acquisition and pro-cessing with emphasis on spectral and wide-dynamicrange imaging, photonic lab-on-a-chip, point-of-care

screening, and biochemical detection applications.Mr. Ho was the recipient of the Postgraduate Scholarship and Undergraduate

Student Research Award from the Natural Sciences and Engineering ResearchCouncil of Canada. He was also awarded the Ontario Graduate Scholarship inScience and Technology, and the Mary H. Beatty Fellowship.

M. Omair Noor received the Hon. B.Sc. degreein Biotechnology specialist program in 2008, andthe M.Sc. degree in bio-analytical chemistry in2010 from the University of Toronto Mississauga,Mississauga, ON, Canada. He is currently workingtowards the Ph.D. degree in the Department ofChemistry at the University of Toronto with theChemical Sensors Group under the supervision ofProf. Ulrich J. Krull.His research interests focus on the integration of

quantum dots and FRET based assays for nucleic aciddetection using microfluidic channels and paper as solid supports.Mr. Noor currently holds an Ontario Graduate Scholarship (OGS) from the

Ontario Ministry of Training, Colleges and Universities (MTCU). He was alsoawarded the Undergraduate Student Research Award (USRA) from the NaturalSciences and Engineering Research Council of Canada (NSERC).

Ulrich J. Krull is appointed as a Professor ofAnalytical Chemistry at the University of Toronto,and holds the endowed AstraZeneca Chair inBiotechnology. His research interests are in theareas of biosensor and diagnostic technologies, andapplications to biotechnology, forensic, clinicaland environmental chemistry. His research workis exploring the use of nanoscale materials andmicrofluidics technologies to build devices fordetection of DNA and RNA targets. Prof. Krull is aneditor for Analytica Chimica Acta, and serves on a

number of scientific advisory boards for industry.

Glenn Gulak (M’83–SM’96) received the Ph.D.degree from the University of Manitoba, Canada,while holding a Natural Sciences and EngineeringResearch Council of Canada Postgraduate Schol-arship. From January 1985 to January 1988 he wasa Research Associate in the Information SystemsLaboratory and the Computer Systems Laboratoryat Stanford University.He is a Professor in the Department of Electrical

and Computer Engineering at the University ofToronto, ON, Canada. He is a registered Professional

Engineer in the Province of Ontario. His present research interests are in theareas of algorithms, circuits, and CMOS system-on-chip implementationsfor digital communication systems and, additionally, in the area of CMOSbiosensors. His current research projects are focused on high-performanceMIMO OFDM implementations and in CMOS biosensors. He has authoredor co-authored more than 100 publications in refereed journal and refereedconference proceedings. He has received numerous teaching awards forundergraduate courses taught in both the Department of Computer Scienceand the Department of Electrical and Computer Engineering at the Universityof Toronto. He held the L. Lau Chair in Electrical and Computer Engineeringfor the 5-year term from 1999 to 2004. From March 2001 to March 2003 hewas the Chief Technical Officer and Senior VP LSI Engineering of a fablesssemiconductor startup headquartered in Irvine, CA, USA. He currently holdsthe Canada Research Chair in Signal Processing Systems and the Edward S.Rogers Sr. Chair in Electrical Engineering.Dr. Gulak has served on the ISSCC Signal Processing Technical Subcom-

mittee from 1990 to 1999, ISSCC Technical Vice-Chair in 2000 and served asthe Technical Program Chair for ISSCC 2001. He received the IEEE Millen-nium Medal in 2001. He served on the Technology Directions Subcommitteefor ISSCC from 2005 to 2008. He currently serves as the Chair of the Publica-tions Committee for the IEEE Solid-State Circuits Society.

Roman Genov (M’96–SM’02) received the B.S.degree (first rank) in electrical engineering fromRochester Institute of Technology, Rochester, NY,USA, in 1996, and the M.S. and Ph.D. degrees inelectrical and computer engineering from JohnsHopkins University, Baltimore, MD, USA, in 1998and 2002, respectively.He has held engineering positions with Atmel Cor-

poration, Columbia, MD, USA, in 1995, and XeroxCorporation, Rochester, NY, USA, in 1996. He was aVisiting Researcher with the Laboratory of Intelligent

Systems, Swiss Federal Institute of Technology (EPFL), Lausanne, Switzer-land, in 1998 and with the Center for Biological and Computational Learning,Massachusetts Institute of Technology, Cambridge, MA, USA, in 1999. He ispresently an Associate Professor with the Department of Electrical and Com-puter Engineering, University of Toronto, Toronto, ON, Canada. His researchinterests include analog and digital VLSI circuits, systems, and algorithms forenergy-efficient signal processing with applications to electrical, chemical, andphotonic sensory information acquisition, biosensor arrays, brain-silicon inter-faces, parallel signal processing, adaptive computing for pattern recognition,and implantable and wearable biomedical electronics.Dr. Genov received the Canadian Institutes of Health Research (CIHR)

Next Generation Award in 2005, the Brian L. Barge Award for excellencein microsystems integration in 2008, the DALSA Corporation Award forexcellence in microsystems innovation in 2006 and 2009, and the Best PaperAward on sensors and Best Student Paper Award, both at the IEEE Interna-tional Symposium on Circuits and Systems in 2009. He is an associate editorof IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, IEEETRANSACTIONS ON CIRCUITS AND SYSTEMS II: EXPRESS BRIEFS, and IEEESignal Processing Letters.

2116 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL…roman/professional/... ·...

Documents

Transcript of 2116 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL…roman/professional/... ·...